Download Embedding Values Into Autonomous Intelligent Systems

The IEEE Global Initiative for Ethical Considerations in Artificial Intelligence and Autonomous Systems
2
Embedding Values Into Autonomous
Intelligent Systems
Society does not have universal standards or guidelines to help embed human norms
or moral values into autonomous intelligent systems (AIS) today. But as these systems
grow to have increasing autonomy to make decisions and manipulate their environment,
it is essential they be designed to adopt, learn, and follow the norms and values of the
community they serve, and to communicate and explain their actions in as transparent
and trustworthy manner possible, given the scenarios in which they function and the
humans who use them.
The conceptual complexities surrounding what “values” are make it currently difficult to
envision AIS that have computational structures directly corresponding to values. However,
it is a realistic goal to embed explicit norms into such systems, because norms can be
considered instructions to act in defined ways in defined contexts. A community’s network
of norms as a whole is likely to reflect the community’s values, and AIS equipped with
such a network would therefore also reflect the community’s values, even if there are
no directly identifiable computational structures that correspond to values.
To address this need, our Committee has broken the broader objective of embedding
values into these systems into three major goals:
1. Identifying the norms and eliciting the values of a specific community affected by AIS.
2. Implementing the norms and values of that community within AIS.
3. Evaluating the alignment and compatibility of those norms and values between the
humans and AIS within that community. This work is licensed under a Creative Commons Attribution-NonCommercial 3.0 United States License.
22
The IEEE Global Initiative for Ethical Considerations in Artificial Intelligence and Autonomous Systems
2
Embedding Values Into Autonomous
Intelligent Systems
Pursuing these three goals represents an iterative process that is contextually sensitive to
the requirements of AIS, their purpose, and their users within a specific community. It is
understood that there will be clashes of values and norms when identifying, implementing,
and evaluating these systems (a state often referred to as “moral overload”). This is why
we advocate for a stakeholder-inclusive approach where systems are designed to provide
transparent signals (such as explanations or inspection capabilities) about the specific
nature of their behavior to the various actors within the community they serve. While this
practice cannot always eliminate the possible data bias present in many machine-learning
algorithms, it is our hope that the proactive inclusion of users and their interaction with AIS
will increase trust in and overall reliability of these systems.
This work is licensed under a Creative Commons Attribution-NonCommercial 3.0 United States License.
23
The IEEE Global Initiative for Ethical Considerations in Artificial Intelligence and Autonomous Systems
2
Embedding Values Into Autonomous Intelligent Systems
Identifying Norms and Values
for Autonomous Intelligent Systems
Issue:
Values to be embedded in
AIS are not universal, but
rather largely specific to user
communities and tasks.
Background
If machines enter human communities as
autonomous agents, then those agents will
be expected to follow the community’s social
and moral norms. A necessary step in enabling
machines to do so is to identify these norms.
Whereas laws are formalized and therefore
relatively easy to identify, social and moral
norms are more difficult to ascertain, as they are
expressed through behavior, language, customs,
cultural symbols, and artifacts. Moreover,
communities (from families to whole nations)
differ to various degrees in the norms they follow.
So embedding norms in AIS requires a clear
delineation of the community in which AIS are
to be deployed. Further, even within the same
community, different types of AIS will demand
different sets of norms. The relevant norms
for self-driving vehicles, for example, will differ
greatly from those for robots used in healthcare.
Candidate Recommendation
We acknowledge that generating a universal
set of norms/values that is applicable for all
autonomous systems is not realistic. Instead,
we recommend to first identify the sets of norms
that AIS need to follow in specific communities
and for specific tasks. Empirical research involving
multiple disciplines and multiple methods should
investigate and document these numerous sets
of norms and make them available for designers
to implement in AIS.
Further Resources
This book describes some of the challenges of
having a one-size-fits-all approach to embedding
human values in autonomous systems: Wallach,
Wendell and Colin Allen. Moral Machines:
Teaching Robots Right from Wrong. Oxford
University Press, 2008.
This work is licensed under a Creative Commons Attribution-NonCommercial 3.0 United States License.
24
The IEEE Global Initiative for Ethical Considerations in Artificial Intelligence and Autonomous Systems
2
Embedding Values Into Autonomous Intelligent Systems
Issue:
Moral overload – AIS are usually
subject to a multiplicity of
norms and values that may
conflict with each other.
Background
An autonomous system is often built with many
constraints and goals in mind. These include
legal requirements, monetary interests, and also
social and moral values. Which constraints should
designers prioritize? If they decide to prioritize
social and moral norms of end users (and other
stakeholders), how would they do that?
Candidate Recommendation
Our recommended best practice is to prioritize
the values that reflect the shared set of values
of the larger stakeholder groups. For example,
a self-driving vehicle’s prioritization of one factor
over another in its decision making will need to
reflect the priority order of values of its target
user population, even if this order is in conflict
with that of an individual designer, manufacturer,
or client. For example, the Common Good
Principlevii could be used as a guideline to
resolve differences in the priority order of
different stakeholder groups.
We also recommend that the priority order
of values considered at the design stage of
autonomous systems have a clear and explicit
rationale. Having an explicitly stated rationale for
value decisions, especially when these values are
in conflict with one another, not only encourages
the designers to reflect on the values being
implemented in the system, but also provides
a grounding and a point of reference for a third
party to understand the thought process of
the designer(s). The Common Good Principle
mentioned above can help formulate such
rationale.
We also acknowledge that, depending on the
autonomous system in question, the priority
order of values can dynamically change from
one context of use to the next, or even within
the same system over time. Approaches such
as interactive machine learning (IML), or direct
questioning and modeling of user responses
can be employed to incorporate user input into
the system. These techniques could be used
to capture changing user values.
Further Resources
•
Markkula Center for Applied Ethics, The
Common Good. Idea of the common good
decision-making was introduced here.
•
Van den Hoven, Jeroen, Engineering and
the Problem of Moral Overload. Science and
Engineering Ethics 18, no. 1 (March 2012):
143-155.
•
One of the places where differences in
human moral decision-making and changes
in priority order of values for autonomous
systems are documented is a series of poll
results published by the Open Roboethics
initiative. In particular, see these poll results
on care robots.
This work is licensed under a Creative Commons Attribution-NonCommercial 3.0 United States License.
25
The IEEE Global Initiative for Ethical Considerations in Artificial Intelligence and Autonomous Systems
2
Embedding Values Into Autonomous Intelligent Systems
Issue:
AIS can have built-in data
or algorithmic biases that
disadvantage members
of certain groups.
Background
Autonomous intelligent systems, compared to
traditional systems, are sometimes discussed
as a new type of species—called the new
ontological category,vii according to literature in
human-robot interaction—because of the manner
in which humans perceive, interact with, and
psychologically respond to them. For example,
numerous studies have documented the way in
which humans willingly follow even the strangest
of requests from a robot, demonstrating the
impact these systems can have on our decisionmaking and behavior (see for example, Robinette,
Paul, Wenchen Li, Robert Allen, Ayanna M.
Howard, and Alan R. Wagner, “Overtrust of
Robots in Emergency Evacuation Scenarios,”viii
2016 ACM/IEEE International Conference on
Human-Robot Interaction). Hence, it is important
to be aware of possible use of the systems for
the purposes of manipulation.
In addition, various aspects of these systems
can be designed to instill bias into other users,
whether intended or not. The sources of bias can
span from the way a system senses the world
(e.g., can the system detect a person missing
an arm or does it assume all humans have two
arms?), to how it processes and responds to
the sensed information (e.g., does the system
respond to people of different ethnicity, gender,
race, differently?), as well as what it looks like.
Details of an interactive autonomous system’s
behavior can have far-reaching consequences,
such as reinforcement of gender, ethnic, and
other biases (see for example, Bolukbasi,
Tolga, Kai-Wei Chang, James Zou, Venkatesh
Saligrama, and Adam Kalai, “Man Is to Computer
Programmer as Woman Is to Homemaker?
Debiasing Word Embeddings,”ix Cornell University
Library, arXiv:1607.06520, July 21, 2016.)
Moreover, while deciding which values and
norms to prioritize, we call for special attention
to the interests of vulnerable and underrepresented populations, such that these user
groups are not exploited or disadvantaged by
(possibly unintended) unethical design. While
traditionally the term vulnerable populations
refers to disadvantaged sub-groups within
human communities—including but not limited to
children, older adults, prisoners, ethnic minorities,
economically disadvantaged, and people with
physical or intellectual disabilities—here we also
include populations who may not be traditionally
considered a member of vulnerable populations,
but may be so in the context of autonomous
intelligent systems. For example, riders in
autonomous vehicles, or factory workers using
a 400-pound high-torque robot, who would not
otherwise be vulnerable under the traditional
definition, become vulnerable in the use contexts
due to the user’s reliance on the system or
physical disadvantage compared to the highpowered machinery.
This work is licensed under a Creative Commons Attribution-NonCommercial 3.0 United States License.
26
The IEEE Global Initiative for Ethical Considerations in Artificial Intelligence and Autonomous Systems
2
Embedding Values Into Autonomous Intelligent Systems
Candidate Recommendation
It is important to acknowledge that it is easy to
have built-in biases in autonomous systems.
For example, a system that depends on face
recognition trained entirely on Caucasian faces
may work incorrectly or not at all on people with
non-Caucasian skin tones or facial structures.
This renders the system to be perceived as
discriminatory, whether it was designed with
such intent or not. These biases can also stem
from the values held by the designer. We can
reduce the incidence of such unintended biases
by being more aware of the potential sources
of these biases. We posit that being aware of
this particular issue and adopting more inclusive
design principles can help with this process.
For example, systems that can sense persons
of different races, ethnicities, genders, ages,
body shapes, or people who use wheelchairs or
prosthetics, etc.
We also highlight that this concern delves into
the domain of ongoing research in human-robot
interaction and human-machine interaction. To
what extent and how do built-in biases change
the course of robot interaction with human users?
What dynamic and longitudinal effect do they
have on the users and the society? How does a
robot’s morphology in different use cases affect
target user groups? These are all open research
questions for which we do not yet have clear
answers. Since there is no clear understanding
of the nature of these biases and their alignment
with human values, we recommend conducting
research and educational efforts to resolve these
open questions and to address these issues in a
participatory way by introducing into the design
process members of the groups who may be
disadvantaged by the system.
In particular, vulnerable populations are often
one of the first users of autonomous systems. In
designing for these populations, we recommend
designers familiarize themselves with relevant
resources specific to the target population. We
also note that a system can have multiple end
users, each of which may demand a conflicting
set of values. We recommend designers be aware
of such conflicts and be transparent in addressing
these conflicting value priorities as suggested
in the above-mentioned issue. AIS are usually
subject to a multiplicity of norms and values that
may conflict with each other.
Therefore, we strongly encourage the inclusion of
intended stakeholders in the entire engineering
process, from design and implementation to
testing and marketing, as advocated for example
in disability studies literature (see “Nothing About
Us Without Us” in the Further Resources below).
A number of institutions have established
connections with communities of a particular
vulnerable population (e.g., University of
Washington’s DO-IT program). However, there
is no one voice that represents all vulnerable
populations. Hence, we recommend designers
and practitioners reach out to communities of
interest and relevant advocacy groups.
We also recommend, especially when designing
for dynamically vulnerable populations, that
designers take on an interdisciplinary approach
and involve relevant experts or advisory group(s)
into the design process. Thus, designers of AIS
should work together with behavioral scientists
This work is licensed under a Creative Commons Attribution-NonCommercial 3.0 United States License.
27
The IEEE Global Initiative for Ethical Considerations in Artificial Intelligence and Autonomous Systems
2
Embedding Values Into Autonomous Intelligent Systems
and members of the target populations
to systematically study population norms,
expectations, concerns, and vulnerabilities. We
also encourage designers to include regulators
and policymakers in this process as well, noting
that shaping regulation and policy is an integral
part of guiding the development and deployment
of autonomous systems in a desirable direction.
•
Federal Trade Commission. Privacy Online:
Fair Information Practices in the Electronic
Marketplace: A Federal Trade Commission
Report to Congress. 2000.
•
Riek, Laurel D. “Robotics Technology in
Mental Health Care.” Artificial Intelligence in
Behavioral Health and Mental Health Care,
(2015): 185-203.
Further Resources
•
Charlton, James I. Nothing About Us Without
Us: Disability Oppression and Empowerment,
University of California Press, 2000.
•
Shivayogi, P. “Vulnerable Population and
Methods for Their Safeguard.” Perspectives
in Clinical Research, January-March (2013):
53-57.
•
Asaro, P. “Will BlackLivesMatter to RoboCop? ”
We Robot, 2016.
•
Riek, L. D. and D. Howard. A Code of Ethics
for the Human-Robot Interaction Profession.
We Robot, 2014.
•
Winfield, A. Robots Should Not Be Gendered
(blog), 2016.
•
Whitby, Blay. “Sometimes It’s Hard to Be
a Robot: A Call for Action on the Ethics of
Abusing Artificial Agents.” Interacting with
Computers 20, no. 3 (2008): 326-333.
This work is licensed under a Creative Commons Attribution-NonCommercial 3.0 United States License.
28
The IEEE Global Initiative for Ethical Considerations in Artificial Intelligence and Autonomous Systems
2
Embedding Values Into Autonomous Intelligent Systems
Embedding Norms and Values
in Autonomous Intelligent Systems
Issue:
Once the relevant sets of
norms (of AIS’s specific role in
a specific community) have been
identified, it is not clear how
such norms should be built into
a computational architecture.
Background
The prospect of developing computer systems
that are sensitive to human norms and values
and factoring these issues into making decisions
in morally or legally significant situations has
intrigued science fiction writers, philosophers,
and computer scientists alike. Modest efforts
to realize this worthy goal in limited or bounded
contexts are already underway. This emerging
field of research goes under many names
including: machine morality, machine ethics,
moral machines, value alignment, computational
ethics, artificial morality, safe AI, and friendly
AI. Basic notions can be found in books such
as Allen, C., and W. Wallach. Moral Machines. x
Teaching Robots Right from Wrong. Oxford
University Press, 2010.
Computers and robots already instantiate values
in their choices and actions, but these values
are programmed or designed by the engineers
that build the systems. Increasingly, autonomous
systems will encounter situations that their
designers cannot anticipate, and will require
algorithmic procedures to select the better of
two or more possible courses of action. Some of
the existing experimental approaches to building
moral machines are top-down. In this sense the
norms, rules, principles, or procedures are used
by the system to evaluate the acceptability of
differing courses of action or as moral standards
or goals to be realized.
Recent breakthroughs in machine learning and
perception will enable researchers to explore
bottom-up approaches—in which the AI system
learns about its context and about human
values—similar to the manner in which a child
slowly learns which forms of behavior are safe
and acceptable. Of course a child can feel
pain and pleasure, empathize with others, and
has other capabilities that AI system cannot
presently imitate. Nevertheless, as research on
autonomous systems progresses, engineers
will explore new ways to either simulate these
capabilities, or build alternative mechanisms that
fulfill similar functions.
This work is licensed under a Creative Commons Attribution-NonCommercial 3.0 United States License.
29
The IEEE Global Initiative for Ethical Considerations in Artificial Intelligence and Autonomous Systems
2
Embedding Values Into Autonomous Intelligent Systems
Candidate Recommendation
Research on this front should be encouraged.
Advances in data collection, sensor technology,
pattern recognition, machine learning, and
integrating different kinds of data sets will enable
creative, new approaches for ensuring that
the actions of AI systems are aligned with the
values of the community in which they operate.
Progress toward building moral machines may
well determine the safety and trustworthiness
of increasingly autonomous AI systems.
•
Anderson, M., and S. Anderson (eds.).
Machine Ethics. Cambridge University Press,
2011.
•
Abney, K., G. Bekey, and P. Patrick. Robot
Ethics: The Ethical and Social Implications
of Robotics. MIT Press, 2011.
•
RC Arkin, P Ulam, AR Wagner, Moral
decision making in autonomous systems:
Enforcement, moral emotions, dignity, trust,
and deception, Proceedings of the IEEE 100
(3), 571-589
Further Resources
•
Allen, C., and W. Wallach. Moral Machines:
Teaching Robots Right from Wrong. Oxford
University Press, 2010.
This work is licensed under a Creative Commons Attribution-NonCommercial 3.0 United States License.
30
The IEEE Global Initiative for Ethical Considerations in Artificial Intelligence and Autonomous Systems
2
Embedding Values Into Autonomous Intelligent Systems
Evaluating the Alignment of Norms
and Values between Humans and AIS
Issue:
Norms implemented in
AIS must be compatible with
the norms in the relevant
community.
Background
If a community’s systems of norms (and their
underlying values) has been identified, and
if this process has successfully guided the
implementation of norms in AIS, then the third
step in value embedding must take place:
rigorous testing and evaluation of the resulting
human-machine interactions regarding these
norms.
An intuitive criterion in these evaluations might
be that the norms embedded in AIS should
correspond closely to the human norms
identified in the community—that is, AIS should
be disposed to behave the same way that
people expect each other to behave. However,
for a given community and a given AIS task and
use context, AIS and humans may not have
identical, but rather compatible, sets of norms.
People will have some unique expectations for
humans that they don’t have for machines (e.g.,
norms governing the expression of emotions, as
long as machines don’t have, or clearly express,
emotions), and people will have some unique
expectations of AIS that they don’t have for
humans (e.g., that the machine will destroy
itself if it can thereby prevent harm to a human).
The norm identification process must document
these structural relations (similarities as well as
differences) between human and AIS norms,
and in evaluating these relations, the goal of
compatibility may be preferred over that of
alignment, which suggests primarily a similarity
structure.
In addition, more concrete criteria must be
developed that indicate the quality of humanmachine interactions, such as human approval
and appreciation of AIS, trust in AIS, adaptability
of AIS to humans users, and human benefits
in the presence or influence of AIS. Evaluation
of these and other criteria must occur both
before broad deployment and throughout the
life cycle of the system. Assessment before
deployment would best take place in systematic
test beds that allow human users (from the
defined community) to engage safely with AIS
(in the defined tasks) and enable assessment of
approval, trust, and related variables. Examples
include the Tokku testing zones in Japan. xi
This work is licensed under a Creative Commons Attribution-NonCommercial 3.0 United States License.
31
The IEEE Global Initiative for Ethical Considerations in Artificial Intelligence and Autonomous Systems
2
Embedding Values Into Autonomous Intelligent Systems
Candidate Recommendation
The success of implementing norms in AIS must
be rigorously evaluated by empirical means,
both before and throughout deployment. Criteria
of such evaluation will include compatibility of
machine norms and human norms (so-called
value alignment or compliance, depending on
the nature of the norms), human approval of
AIS, and trust in AIS, among others. Multiple
disciplines and methods should contribute to
developing and conducting such evaluation, such
as extensive tests (including adversarial ones),
explanation capabilities to reconstruct AIS inner
functioning, natural language dialog between
AIS and humans (including deep question
answering), and context awareness and memory
(to handle repeated evaluations).
Issue:
Achieving a correct level of trust
between humans and AIS.
Background
Development of autonomous systems that are
worthy of our trust is challenged due to the
current lack of transparency and verifiability
regarding these systems for users. For this issue,
we explore two levels at which transparency and
verifiability are useful and often necessary. A first
level of transparency relates to the information
conveyed to the user while an autonomous
system interacts with the user. A second level
has to do with the possibility to evaluate
the system as a whole by a third party (e.g.,
regulators, society at large, and post-accident
investigators).
In the first level, consider for example the case
of robots built to interact with people. The robots
should be designed to be able to communicate
what they are about to perform and why as the
actions unfold. This is important in establishing
an appropriate level of trust with the user.
While a system that a user does not trust may
never be used, a system that is overly trusted
can negatively affect the user as well based
on the perception of the particular system or
similar types of systems by the society. Unlike
humans who naturally use verbal and nonverbal
behaviors to convey trust-based information to
those around them, the mode and the content
of communicative behaviors toward or from an
autonomous system are features that would be
absent if not for the explicit implementation by
the designers. Designing systems that are worthy
of our trust necessarily includes making these
explicit design decisions. As with people, trust
is built over time, through repeated interactions,
so AIS must be equipped with context awareness
and memory capabilities.
Candidate Recommendation
Transparency and verifiability are necessary
for building trust in AIS. We recommend that
AIS come equipped with a module assuring
some level of transparency and verifiability.
Technological solutions to address the issue of
transparency and instilling the right level of trust
in the users is an open area of research. Trust
This work is licensed under a Creative Commons Attribution-NonCommercial 3.0 United States License.
32
The IEEE Global Initiative for Ethical Considerations in Artificial Intelligence and Autonomous Systems
2
Embedding Values Into Autonomous Intelligent Systems
is also a dynamic variable in human-machine
interaction; the level of trust a user may have
with a system tends to change over time.
Coupled with the dynamic nature of trust in
autonomous systems is our known tendency
to overly trust technology beyond its capabilities.
With systems that have been commercialized,
for example, users often assume a minimum
level of reliability and trustworthiness of the
system from the onset.
in not only ineffective interaction with the user—
introducing a point of miscommunication, for
example—but also risk the safety and wellbeing
of the user and others. This also makes it more
challenging for a user to diagnose the reasons
why a system may be behaving in a certain way,
and to detect when malfunctions occur.
Hence, even when a system is delivered with
a written disclaimer outlining its conditions
of use, it is often naïve to assume that the
disclaimer alone can protect the interests of
both the manufacturer/developer and users.
In addition to communicating the limitations
and capabilities of the system to the users,
we recommend autonomous systems to be
designed with features that prevent users from
operating the system outside a known, safe, and
appropriate range of conditions of use, including
conditions that depend on user behavior. We also
recommend evaluation of the system’s design
with the user’s perception of their role in mind
(e.g., operator versus user of the system), such
that the system’s interaction with the user is in
alignment with the role that is expected of the
user.
Issue:
Third-party evaluation of AIS’s
value alignment.
In addition, one can design communicative
and behavioral features of a system to serve
as interactive real-time disclaimers, such that
the user is informed of significant changes to
the system’s level of confidence on a proposed
solution for the task to be performed, which can
change from one moment or situation to the
next. Systems that lack such features can result
Background
The second level of transparency, as stated
above, is needed to evaluate a system as a whole
by a third party (e.g., regulators, society at large,
and post-accident investigators).
In this second category, there are concerns
regarding the increasing number of autonomous
systems that rely on, or include, AI/machinelearning techniques inherently lacking
transparency and verifiability. Discussions on this
topic include: the nature and possible bias of the
data sets used to train a machine-learning system
that is often not accessible by the public, details
of the algorithm used to create the final product,
the specifications on the final product’s efficacy
and performance, and the need to consider the
scenario where AIS will be used when evaluating
their adherence to relevant human values. While
acknowledging the usefulness and potential for
This work is licensed under a Creative Commons Attribution-NonCommercial 3.0 United States License.
33
The IEEE Global Initiative for Ethical Considerations in Artificial Intelligence and Autonomous Systems
2
Embedding Values Into Autonomous Intelligent Systems
these systems, it is a serious concern that even
the designers and programmers involved cannot
verify or guarantee reliability, efficacy, and value
alignment of the final system. A further problem
is that there is no agreed-upon method, process,
or standards for validating and certifying the
adherence of AIS to desired human norms and
values.
Candidate Recommendation
With regards to our concern on the transparency
between a system as a whole and its evaluator
(e.g., regulator), we recommend that designers
and developers alike document changes to the
systems in their daily practice. A system with
the highest level of traceability would contain
a black-box-like module such as those used in
the airline industry, that logs and helps diagnose
all changes and behaviors of the system. Such
practice, while it does not fully address the
need for transparency of a number of popular
machine-learning approaches, allows one to trace
back to the sources of problems that may occur
and provide a mechanism with which a faulty
behavior of a system can be diagnosed.
As more human decision-making is delegated
to autonomous systems, we expect there to be
an increasing need for rationale and explanation
as to how the decision was reached by the
algorithm. In this respect, a relevant regulation
is the European Union’s new General Data
Protection Regulation (GDPR) xiv, adopted on
April 2016 and scheduled to take effect in 2018.
The GDPR states that, in regards to automated
decisions based on personal data, individuals
have a right to “an explanation of the [algorithmic]
decision reached after such assessment and to
challenge the decision.” While the development
of an algorithm that is able to explain its behavior
is an open research topic, there are algorithms
that are more transparent than others, such as
logic-based AI that provide more transparency
than machine-learning AI, and more coherence
between the output behavior of a system and
its inner functioning. Winfield, Blum, and Liu’s
work on consequence enginexv, for example,
utilizes a simulator to predict and evaluate the
consequences of an artificial agent’s possible
next actions in order to decide the right course
of action, making the agent’s decision-making
process easy to examine and validate. In
the absence of an adequate alternative, it is
imperative that designers be aware of the need
for transparency and strive to increase it in the
algorithms they design and implement into
autonomous systems.
We also recommend that regulators define,
together with users, developers, and designers,
a minimum level of value alignment and
compliance, and suitable capabilities for this
to be checked by a third party, in order for
AIS to be deployed.
Finally, we recommend to define criteria to define
AIS as trustworthy. These criteria will depend on
a machine’s expected tasks and context of use,
as well as the users’ vulnerabilities (we expect
that more-vulnerable-user categories will require
more stringent criteria).
This work is licensed under a Creative Commons Attribution-NonCommercial 3.0 United States License.
34
The IEEE Global Initiative for Ethical Considerations in Artificial Intelligence and Autonomous Systems
2
Embedding Values Into Autonomous Intelligent Systems
Further Resources
•
Goodman, B., and S. Flaxman, “European
Union Regulations on Algorithmic DecisionMaking and a ‘Right to Explanation’,”
Cornell University Library, arXiv: 1606.08813,
August 31, 2016.
•
Winfield, A. F. T., C. Blum, and W.
Liu, “Towards an Ethical Robot: Internal
Models, Consequences and Ethical Action
Selection,” Advances in Autonomous
Robotics Systems, Lecture Notes in Computer
Science Volume 8717, (2014): 85-96. Eds.
Mistry M, Leonardis A, Witkowski M and
Melhuish C, Springer, 2014.
This work is licensed under a Creative Commons Attribution-NonCommercial 3.0 United States License.
35